Most AI tutoring tools on the market today are essentially just advanced chat boxes with a handful of pre-built templates—chat, quizzes, and notes all operate in silos, memory is locked away in a black box, and personalization relies entirely on prompt engineering.

DeepTutor v1.5, open-sourced by the Data Intelligence Lab at the University of Hong Kong, takes a fundamentally different approach. Its core design principle is straightforward: **tutoring should be a continuous data cycle, not a disconnected pile of features**.


DeepTutor positions itself as an "agent-native learning workspace"—instruction, problem solving, question generation, research, visualization, and mastery-based practice all run on the same single agent loop. When you switch between tasks, you're only switching your target, not the underlying engine, and your full context stays with you.

### Six Key Features, Each Solving a Real Pain Point

**1. Unified Runtime**

Chat, Quiz, Research, Visualize, Solve, and Mastery Path all share the same agent loop. No more copying and pasting context between different tools—your memory and knowledge base automatically carry over when you switch modes.

**2. Connected Learning Context**

Knowledge bases, books, Co-Writer drafts, notebooks, question banks, role settings, memory—none of these are isolated files in DeepTutor. They are shared resources accessible across every workflow.

**3. Sub-agents and Partners**

You can spin up Claude Code, Codex, or other partners at any point during a conversation, import their full conversation history, and even let IM bots across 15 platforms (Telegram, Slack, WeChat, and more) share the same unified brain.

**4. Multi-Engine Knowledge**

It supports a wide range of retrieval engines including LlamaIndex, PageIndex, GraphRAG, LightRAG, and Obsidian. Document parsing is pluggable, indexes come with built-in version control, and rebuilding indexes won't corrupt your existing work.

**5. Extensible Tools and Skills**

It comes with built-in tools, MCP servers, and image/video/voice generation models, plus installable skill packages from the EduHub community. All packages go through a security gate during installation: it checks for safety verdicts, defends against zip-slip attacks, and strips forced injection markers.

**6. Inspectable Memory**

It features a three-layer memory system: L1 raw event traces, L2 surface-level summaries, and L3 cross-context synthesis. Every claim can be traced back to its original source, and memory graph visualization lets you see exactly how personalization is built. It's not a black box—it's a transparent workspace.

### Installation and Ecosystem

DeepTutor is fully open-source and offers four installation paths: PyPI, Docker, source code, and pure CLI. The recommended installation method is:

```bash

mkdir my-deeptutor && cd my-deeptutor

pip install -U deeptutor

deeptutor init

deeptutor start

```

Then just open `http://127.0.0.1:3782` in your browser to start using it.

The project also comes with a dedicated EduHub skill community, full documentation at deeptutor.info, and an arXiv paper (2604.26962). Since its release in late 2025, it has already gained 25.2k stars on GitHub.

### A Quick Observation

DeepTutor's design brings up an old question: Why do education apps keep piling on more features, but learning outcomes never actually get better? The answer probably is that there is no connected data cycle between features. DeepTutor builds a closed loop out of memory, behavior, and feedback, and makes this loop fully transparent and auditable. This kind of "inspectable personalization" might get us much closer to true teaching tailored to each individual student than any fancy new model.

Of course, the project is still young. But it's heading in the right direction.

---

- GitHub: [https://github.com/HKUDS/DeepTutor](https://github.com/HKUDS/DeepTutor)

- Official Website: [https://deeptutor.info](https://deeptutor.info)

- Paper: [https://arxiv.org/abs/2604.26962](https://arxiv.org/abs/2604.26962)